GraphVLA V7 — step 31,527
Model and full four-rank DeepSpeed ZeRO-2 resume state, saved during an explicit graceful pause. V7 uses mask-coverage-weighted visual patch pooling. Requires matching custom V7 code; this is not an AutoModel checkpoint. No evaluation performance claim is made by this upload.
Evaluation
Download checkpoints/model.pt, config.yaml and dataset_statistics.json. Use the V7 inference server with the matching perception dependencies. Training data and optimizer files are not needed for evaluation.
Continuing training
Training cache: https://huggingface.co/datasets/Kensag/GraphVLA-V7-Training-Cache This cache is required for continuing the original training sequence. It is not duplicated here. Verify its manifest and hashes before reuse.
Copy resume/step_031527 into a new run directory's step_031527; place the model from checkpoints/model.pt there as model.pt (a local hard link is sufficient). Copy resume/ds_checkpoints/step_031527 to that run's ds_checkpoints/step_031527. Create latest.json with {"step":31527,"directory":"step_031527"}. Use the matching V7 trainer and resume with four ranks, batch 24 per rank, BF16 model, FP32 master weights/moments, ZeRO-2, BPTT=1, original 80k schedule. Do not reset the optimizer/scheduler or use the fresh-start-only launch guard.
This is not a turnkey environment: custom source and all upstream weights, dependencies and relevant permissions must be obtained separately. The V7 source is not made public by this model upload. Full resume includes optimizer, scheduler, RNG, stream progress and graph memory. Metadata retains original machine paths and episode identifiers; relocate paths and re-run manifest validation explicitly, never bypass the trainer's integrity checks. Exact numerical continuation across different hardware is not guaranteed. Upstream model and dataset terms continue to apply; no new license is granted. See SHA256.json for integrity of all binary/config artifacts.
- Downloads last month
- 31